International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
are images with evident texture of classes, designed with use of 
Brodatz textures (Figure 2). 
  
  
Figure 2. Model image with evident textures of classes 
4.2 Nonparametric density estimation 
The efficiency investigation of density estimation algorithms is 
performed for the determination of the computational cost 
(Figure 3 a)) and the accuracy (Figure 3 b)) of classification, on 
the example of seven bands image of type as shown in Figure 1. 
10000 ] 
pan a N 
1000 | tir 
Öö ! E ud —«PáÁ— | 
o [d 
0 wena 2 
d 
E 3 
= e—— —9— —9——9—— 9 — —À 
10 5 
Rom dn 
30 350 735 1120 1500 
sample size, pixels 
  
a) 
08 
075 
07 
5 mag 1 
© 065 sr 2 
= 
© 
055 4 
05 
045 
to 
o 
350 735 1120 1500 
sample size, pixels 
b) 
Figure 3. Computational cost and accuracy of classification 
algorithms used 
In Figure 3 the following symbols are defined: 1 — using of 
traditional maximum likelihood classification, 2 — using of 
ordinary RP algorithm based upon (2), 3 — using ol ordinary k- 
NN algorithm based upon (3) and (4), 4 — using the proposed 
original statistical nonparametric algorithm. 
Figure 3 a) shows the original algorithm provides 
computational performance increase in dozens times compared 
to traditional nonparametric density estimation algorithms. At 
that the performance of original algorithm is increasing together 
with the increasing of sample size. 
At the same time Figure 3 b) shows the accuracy of the 
proposed algorithm is almost the same as the accuracy of 
traditional nonparametric algorithms, and also it should be 
noted the accuracy with use of parametric algorithm of density 
estimation is inappropriate low. It proofs necessity of 
developing nonparametric algorithms that are invariant to the 
distribution in a sample. 
4.3 Classification with spatial features 
The important element of the advanced interpretation is the 
classification of RS images with use of texture features of the 
classes that is needed for high accuracy interpretation. In the 
framework of the developing approach it is proposed to define 
prior probabilities of classes by either statistical or ANN ways 
and each way takes into account the spatial features of classes. 
The efficiency investigation of some algorithms with use of 
different types of model images is conducted. The purpose of 
the research is to define how the proposed ways of forming 
spatial feature space consider texture information about classes. 
1 
0.8 
[nk 
g 06 02 
3 cas 
s 04 04 
0.2 aii 
0 
type 1a type 1b type 2a type 2b 
models 
Figure 4. The classification accuracy of statistical and ANN 
algorithms with different types of model images 
used 
The Figure 4 shows the investigation results of accuracy 
classification for different ways of forming feature space. 
In Figure 4 the following symbols are defined: 1 — traditional 
maximum likelihood classification, 2 — ordinary RP algorithm 
based upon (2), 3 — ordinary k-NN algorithm based upon (3) 
and (4) 4 — the proposed ANN original algorithm using 
context-spectral way of forming feature space, 5 — the proposed 
nonparametric algorithm using Haralick texture characteristics 
(Haralick R.M. & Joo H. A, 1986). The following types of 
model images on the abscissa axis are scaled: typela and 
typelb — the first type model images (Figure 1) with three and 
seven bands consequently; type2a and type2b — the second type 
model images (Figure 2) with only one and six bands 
consequently. 
Figure 4 presents the property to get more accurate results of 
the ANN classification with context-spectral forming feature 
space due to considering texture features in multispectral 
images. The noticeable effect of that in case of ANN 
Inter 
class 
almo 
stror 
At t 
prop 
upon 
nonp 
Besi 
mutl 
conc 
majc 
bush 
Figu 
Figu 
200€ 
obta 
max 
tradi 
c) b 
enha 
Figu 
a les